Document Forensic Pipeline

Two-branch fraud detection for identity documents: face/photo tampering + text field tampering.

Public leaderboard: 0.294 (FREUID score, lower is better)

Pipeline

1. MediaPipe BlazeFace — face detection (2ms/image, CPU)
2. DINOv2-small — portrait crop → photo fake probability (1.2ms/face, GPU)
3. PaddleOCR PP-OCRv6 — text field detection (35ms/image, GPU)
4. Forensic panels — each field crop → 224×1008 six-view composite (grayscale, ink mask, L residual, chroma, texture, edge)
5. ConvNeXt-Base — panel → per-field tamper probability (1.5ms/field, GPU)
6. Score — fraud_score = max(photo_prob, max(field_probs))

Total inference: ~6h for 135k documents on A100.

Repository Structure

├── prepare_submission.py           # Docker entrypoint
├── Dockerfile                      # GPU Docker image (CUDA 12.8)
├── requirements.txt                # Python dependencies
├── technical_report.pdf            # Compiled technical report
├── LICENSE                         # MIT
│
├── inference/
│   └── run_inference.py            # Full inference pipeline
│
├── models/
│   ├── models.py                   # DINOv3Classifier, DINOClassifier
│   └── weights/                    # Via Git LFS (run `git lfs pull` after clone)
│
├── train/
│   ├── how_to_train.md              # Complete training guide
│   ├── train_text.py               # Train ConvNeXt-B text tamper detector
│   ├── train_face.py               # Train DINOv2-small face detector
│   ├── precache_boxes.py           # Pre-run PaddleOCR on training images
│   ├── train_labels.csv            # Competition training labels
│   ├── annotations/                # Subtask + field-level annotations
│   └── scripts/                    # Shared modules (panels, data, synth, metric)
│
├── technical_report/               # LaTeX source + figures
│   ├── technical_report.tex
│   ├── references.bib
│   └── figures/
│
└── REPLY_TEMPLATE.txt              # Kaggle discussion reply template

Model Weights

Weights are stored via Git LFS. After cloning, pull the actual files:

git clone https://github.com/akhil838/doc-forensic-pipeline.git
cd doc-forensic-pipeline
git lfs pull

Also available on HuggingFace: https://huggingface.co/akhil838/doc-forensic-models-v1

Docker Build & Run

# Build
docker build -t freuid-repro .

# Run (no network)
docker run --rm --gpus all --network none \
  -v /path/to/test/images:/data:ro \
  -v $(pwd)/out:/submissions \
  freuid-repro:latest

Output: /submissions/submission.csv with columns id,label.

Local Inference (without Docker)

python inference/run_inference.py \
  --image-dir /path/to/images \
  --model-dir models/weights \
  --output submission.csv \
  --device cuda

Training

See train/how_to_train.md for complete instructions.

Hardware

• Training: Apple M3 Max (16-core CPU, 40-core GPU, 64GB unified RAM, MPS)
• Inference: NVIDIA A100-SXM4-40GB, 16-core CPU, 120GB RAM, 130GB disk (~6h for 135k docs)

External Resources

No external training data beyond the FREUID competition dataset.

Contact

Akhil Kosuri — kosuriakhil19@gmail.com